Skip to content
AnthropicAnthropic

An initiative to secure the world's software | Project Glasswing

Project Glasswing is a new initiative that brings together Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks in an effort to secure the world’s most critical software. We formed Project Glasswing because of capabilities we’ve observed in a new frontier model trained by Anthropic that we believe could reshape cybersecurity. Claude Mythos Preview is a general-purpose, unreleased frontier model that reveals a stark fact: AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities. Read more: https://anthropic.com/glasswing

Apr 7, 20265mWatch on YouTube ↗

CHAPTERS

  1. Why everyday software bugs become a global problem

    The video opens by contrasting how most users rarely notice bugs with how developers face them constantly. It frames the real danger: occasional vulnerabilities that cause severe, widespread impact when they occur in shared components.

  2. The traditional vulnerability cycle is slow and costly

    The discussion explains that discovering and patching vulnerabilities has historically been resource-intensive. This sets up why new approaches are needed as the software ecosystem grows more complex and interconnected.

  3. LLMs raise the stakes for both defenders and attackers

    The video highlights a dual-use shift: models that can write strong code can also find and exploit bugs effectively. This creates a cybersecurity “bar-raising” moment where both defensive and offensive capabilities accelerate.

  4. Claude Mythos Preview: a step-change in cyber capability

    Anthropic introduces Claude Mythos Preview and notes it showed meaningful cybersecurity gains early on. Although not trained specifically for cybersecurity, improved coding ability translated into stronger security-research performance.

  5. Near-professional bug-finding performance—and exploit chaining

    The model is described as roughly as good as a professional human at identifying bugs, with an additional advantage: chaining multiple smaller vulnerabilities into sophisticated exploit paths. Its autonomy supports long-range investigative workflows similar to a human researcher’s day.

  6. Why it won’t be widely released: managing dual-use risk

    Because these capabilities could cause harm if broadly accessible, Anthropic states it will not release the model widely. The video emphasizes that even stronger models will arrive, making risk response planning urgent.

  7. Project Glasswing: targeted partnerships to harden critical code

    Anthropic launches Project Glasswing to put advanced models into the hands of organizations maintaining critical software. The goal is to reduce risk by giving defenders a head start before such tools become broadly available.

  8. Early results: vulnerabilities found across major platforms

    Working with partners, the team reports finding vulnerabilities across many major platforms at a dramatically increased pace. They prioritize scanning foundational open-source systems that underpin internet infrastructure.

  9. Case studies: OpenBSD and Linux vulnerability discoveries and patching

    The video gives concrete examples: an OpenBSD bug present for 27 years enabling remote crashes, and multiple Linux vulnerabilities enabling privilege escalation from unprivileged users. Findings are responsibly disclosed to maintainers, who patch and deploy fixes.

  10. Empowering maintainers: AI as an invaluable defensive tool

    The narrative underscores how difficult it is to maintain critical software and how valuable it is to catch flaws before exploitation. The model is positioned as a force multiplier for maintainers’ limited time and resources.

  11. Government and cross-industry coordination for systemic security

    Anthropic describes outreach to U.S. government officials to assess risks and strengthen defenses collaboratively. The broader message is that software underpins modern life, making cybersecurity a societal security imperative.

  12. A long-term effort to make global software safer

    The video closes by setting expectations: improving the security of global software will take months or years. The intended outcome is measurable reduction in risk to customer data, financial transactions, and critical infrastructure.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome